Data preprocessing

Copy train and test data to torch format

Train data

Test data

Class labels

Torch dataset

Torch dataloader

Resnet-18

Weight decay was added to prevent overfitting

Save model for the future

Explanations

Lime explanations with different masks

Best results are obtainded with segmentation algorithms:

Explanations, that used these masks, show bigger attribuiton of superpixels representing signs. Generally they provide satysfing results.

Missclassified

First, get the missclassified observations from test set.

All missed predictions with images and labels:

Explanation for missed predictions

There are no visible differences in explanations with target set to label or predicted label. Most of attribution is connected to the area of the sign.